AI Patent Eligibility Claims: How Are They Evaluated in 2026?

What Do Courts and the USPTO Require of AI Patent Eligibility Claims in 2026?

As of September 24, 2026, an AI patent claim is not excluded from protection merely because it involves artificial intelligence. Under 35 U.S.C. § 101, software-based claims are tested against the two-step framework from Mayo and Alice: the first step asks whether the claim recites a judicial exception such as a mathematical concept, a certain method of organizing human activity, or a mental process; the second step asks whether the claim integrates that exception into a practical application or supplies additional elements that amount to more than the exception itself. The USPTO's January 7, 2019 subject matter eligibility guidance, updated for AI on July 17, 2024, groups abstract ideas into those three categories and asks examiners and applicants to look for an inventive concept. Because Congress has not created an AI-specific exception and the Supreme Court has not revisited Alice, every AI eligibility dispute is resolved by applying that framework to the specific claim language, the disclosed technical improvement, and the evidence in the file. Claims that recite a particular technical improvement to the functioning of a computer tend to survive; claims that merely direct a generic computer to perform math or mental steps tend not to.

Also worth reading: What Does the 2026 USPTO AI Patent Eligibility Guidance Actually Change for Applicants? · What Are the Emerging AI Patent Eligibility Trends Heading Into 2027? · How Do Patent Examiners Evaluate Subject Matter Eligibility for Machine Learning Inventions Under Current 2026 Guidelines?

Eligibility is evaluated claim by claim, not by the commercial value of the product or the sophistication of the model. An examiner can find one claim eligible and a closely related claim in the same application ineligible, which is why practitioners pair broad and narrow claims carefully. Recent commentary from IPWatchdog, JD Supra, and law-firm AI practices in 2025–2026 reports higher Section 101 invalidation rates for AI patents than for the broader patent population, reflecting the fact that courts are being asked to test AI claims more often and examiners are applying the 2019 and 2024 guidance more aggressively. Those reported rates vary widely by dataset, technology category, and time period, so a single percentage should be treated as a warning signal rather than a forecast for any particular application. The practical takeaway is that drafting, prosecution, and review strategy now matter as much as the underlying invention.

Why AI Claims Fail Eligibility After Alice: Generic Models and Abstract Data Processing

The recurring failure mode is a claim that recites a machine-learning pipeline in abstract terms: gather data, train a model, generate a prediction, and output a result. The Federal Circuit held in Electric Power Group v. Alstom (2016) and SAP America v. InvestPic (2018) that collecting, analyzing, and displaying information is an abstract idea, and that running the analysis on a computer does not by itself supply the additional elements needed at Alice step two. In Recentive Analytics v. Fox Corp. (2024), the court held that claims to using machine learning to predict an event in order to reduce logistical costs recited an abstract mathematical concept and provided no inventive concept, rejecting the applicant's attempt to map each claim element one-to-one to a mathematical step. Commentators, including Bloomberg Law and firms such as Morgan Lewis, describe that decision as a template for the way generic apply-machine-learning-to-a-new-dataset claims are being rejected, including in specialized domains such as dental imaging.

The reason these claims fail is structural rather than technical. A neural network's training and inference steps are mathematical, and the data relationships they capture are mental in the sense of organizing information, so a claim that stops at those steps is reciting the judicial exception itself. By contrast, in Yu v. Apple (Fed. Cir. 2021), the Federal Circuit found eligible claims reciting a trained model applied to camera images in a specific way to change how the camera captures and processes subsequent images, and in McRO v. Bandai Namco Games (2016), it found eligible claims using specific rules to automate lip synchronization in animation software. The difference is what the claim requires beyond the math: a specific change to computer operation, a specific control scheme, or a specific application tied to the computer's functioning. Drafters who want to survive Alice should therefore show, in the claim itself, what is different about the way the computer operates, not just that the output is useful to a person.

What Changed in 2024–2026: USPTO AI Guidance, SMED Evidence, and Scrutiny

The most visible development for AI patent eligibility claims is the USPTO's July 17, 2024 update to its subject matter eligibility guidance, which added AI-related examples to the 2019 framework. The update reiterates the two-step test and asks whether a claim reciting AI improves the functioning of a computer or otherwise solves a technical problem, and its examples distinguish claims specifying a particular model architecture, training regime, or inference pipeline tied to a technical result from claims reciting AI in purely functional terms. For applicants, that means the specification should expressly connect the model and training to a technical improvement, because the USPTO evaluates whether the claim integrates the exception into a practical application using the claim language read in light of the written description. The 2019 guidance also encourages applicants to structure claims so that one form recites the abstract idea and another recites the inventive concept, a drafting choice that gives examiners and courts a clear path to eligibility.

The USPTO also clarified how evidence can be used in these debates. Commentary from Reed Smith explains that applicants may submit evidence under 37 C.F.R. § 1.132, such as comparative examples or benchmark data, to show that the claimed AI method produces a technical effect not previously known, a practice often called SMED evidence in this context. That evidence can support an argument that additional claim elements supply more than the exception, but it cannot expand the claim beyond what the original disclosure reasonably supports, so the specification still has to do the work. Separately, the USPTO's February 2023 inventorship guidance, which requires a natural person to be named as the inventor, is sometimes confused with eligibility, but it governs who may be listed on the application rather than whether the subject matter is patentable. The combination of the 2024 update, SMED practice, and the heightened invalidation rates reported by IPWatchdog and others explains why 2025–2026 AI prosecution is described as a gauntlet rather than a formality.

What Makes an AI Claim Eligible: Technical Improvement, Particularized Application, Evidence

The strongest eligible AI claims combine three things: a recited technical improvement to computer operation, a particularized application of the model or algorithm, and evidence that the improvement is real. The Federal Circuit's computer-functionality cases, Enfish v. Microsoft (2016) and DDR Holdings v. Hotels.com (2014), stand for the idea that a claim improving the way the computer itself works can reach step two without resort to field-of-use or amount-of-data arguments. In Bascom v. AT&T Mobility (2016), the court approved an eligible claim built from specific ordered steps that arranged data in a nonconventional way, and in Thales Visionix v. United States (2017), it accepted claims reciting a specific sensor-fusion technique. Applied to AI, the analogue is a claim that recites how the model, training data, loss function, or inference schedule changes the computer's memory use, latency, accuracy, or reliability in a defined computing environment, rather than merely reciting that a model exists and produces a result. Commentary from Dykema, Crowell & Moring, and Sterne Keller's AI practice points the same way: eligibility is earned by reciting the technical improvement in the claim, not by describing it only in the specification.

Evidence plays a supporting role. Experimental data, comparison against conventional methods, and proof that the claimed pipeline solves a problem arising in the computer's operation can be offered through declarations and argued at step two. But evidence is not a substitute for claim language, because a court reads the claim as written, and a specification passage not reflected in any claim does not save an abstract claim. There is also a boundary between sections: eligibility under § 101 asks whether the exception is integrated or more is supplied, while §§ 102 and 103 ask whether the claim is new and nonobvious, so a claim can clear one test and fail the other. The USPTO's practice is to apply subject matter eligibility first in examination, so a strong eligibility argument does not substitute for a strong inventive-step position. In short, an AI claim is most likely to survive when a reviewer can point to a specific sentence in the claim explaining what the computer does differently because of the claimed invention.

Practical Steps: Drafting, Prosecution, and Review of AI Claims

The first practical step is to draft claim one around the technical result, not around the model's name. A usable claim recites the specific problem in the computing environment, the specific data and preprocessing steps, the specific model or optimization technique, and the specific change in computer operation or control, while purely functional phrases like configured to determine or analyzing to predict are replaced with algorithmic or structural detail that can be mapped to the specification. Applicants should also prepare fall-back claims that cover both the particularized inventive concept and, where appropriate, a broader recitation of the mathematical idea, consistent with the structure the 2019 guidance encourages. Before filing, an Alice screen is worth running: identify the abstract idea, list the additional elements, and ask in writing whether those elements improve the functioning of the computer or are tied to a technical problem, because that exercise predicts most examiner rejections. The specification should be revised to support any limitation expected to be added later, since amendments that are not supported under § 112 can fail even when they cure § 101.

During prosecution, the standard playbook is an early interview followed by reasoned arguments and, if needed, a Rule 132 declaration with comparative data. Replies should map each claim element to a technical improvement rather than to a field of use, and should distinguish cases like Enfish and Yu from cases like Electric Power Group and Recentive, which is the comparison examiners and examiners' reviewers expect to see. Because the USPTO applies eligibility first, an applicant who fixes eligibility by adding detail should confirm that the added detail does not create an obviousness or written-description problem, a trade-off that is easy to miss under deadline pressure. For portfolio owners, the equivalent step is a claim-by-claim audit that scores each independent claim for recited technical improvement, particularized application, and support, and flags any claim whose eligibility depends on facts appearing only in the summary. Audit findings should be triaged by expiration date, revenue contribution, and enforcement likelihood, since amending a low-value claim may not justify the prosecution and continuation costs.

Comparison: USPTO Examination, Federal Circuit Litigation, and PTAB Review

AI eligibility challenges arrive through three main channels, and each has different timing, cost, and practical consequences. The table below compares those channels using the framework most practitioners use for AI claims; figures are approximate, and current USPTO and PTAB fee schedules should be checked before filing any request.

Because PGR is the only post-grant route that includes § 101, an owner facing an infringement complaint must decide within one year whether to seek it, while an accused infringer facing a complaint generally has an IPR option, but IPR cannot decide § 101 issues. That timing difference is why AI patent review treats the first year after issuance or service of a complaint as the decisive window, and why eligibility weaknesses in a claim set are worth auditing before anyone files suit.

Common Mistakes in AI Patent Eligibility Claims

One common mistake is treating every AI claim as automatically eligible because the invention is technological, when the test instead asks whether the claim recites a specific improvement to the computer. Another is the opposite error: drafting claims so narrowly around a particular model or dataset that the claim is eligible but trivially designed around, which undermines both the value and the enforceability of the patent. A third error is relying on the specification's narrative of advantages without reciting those advantages in the claim, as the court did in Recentive, which looked at the claim elements rather than at the applicant's characterization of them. Practitioners also often conflate three different USPTO actions: the February 2023 AI inventorship rule, which concerns who may be named as the inventor; the 2019 and 2024 eligibility guidance, which concern the subject matter test; and SMED evidence practice, which concerns proof. Confusing them leads to arguments that do not answer the actual rejection.

Timing and amendment discipline produce the rest of the failures. An applicant who amends to overcome a § 101 rejection by adding matter that was never described can trade an eligibility problem for a § 112 written-description problem, and an applicant who amends without a declaration may lose the argument that the added elements are genuinely nonconventional. Portfolio owners make a parallel mistake by waiting until a demand letter arrives, because the one-year windows for PGR and IPR mean a post-issuance attack cannot be planned after the fact. It is also a mistake to assume that favorable USPTO guidance binds the Federal Circuit; the guidance is persuasive, while the court's Alice decisions are controlling. A disciplined review treats each independent claim as a document carrying a burden of proof, and requires the claim itself to carry the technical story.

When to Act, What It Costs, and When a Trade Secret May Be Better

Timing and budget matter as much as drafting. As a general rule, applicants should file before any public disclosure, sale, or offer for sale, because although the United States gives a one-year grace period to the inventor's own disclosure, many foreign rights are lost after a public disclosure, and the Paris Convention generally requires foreign filing within 12 months of the first filing, with national-phase entry typically at 30 months. A provisional application is the cheapest way to buy time, and under the 2025 USPTO fee schedule provisional and utility filing fees range from roughly $160 for a micro entity to about $1,300–$1,600 for a large entity, with issue fees in the $1,000–$1,400 range; the current fee schedule should be checked before relying on those numbers. Attorney costs dominate: a straightforward utility prosecution often runs 250–500 hours at rates of roughly $300–$900 per hour, so an AI-heavy application with intensive eligibility arguments commonly costs from about $150,000 to $400,000 to issue, while a PTAB proceeding adds tens of thousands of dollars in fees and more in counsel time.

Sometimes the better choice is not to prosecute a patent claim at all. A trade secret costs nothing in official fees and can last indefinitely, which makes it attractive where model weights, training recipes, or process know-how are hard to detect by reverse engineering and the product is difficult to monitor. A patent is the better choice where the technology can be inferred from a shipped product, where the team needs defensive publication and priority rights, or where licensing or litigation is expected. A middle path is to file a provisional, control disclosure carefully, and decide at the 12-month mark whether to proceed, keeping in mind that a provisional gives priority but no examination, so it does not test eligibility. Given the heightened invalidation rates reported for AI patents, a reasoned cost-benefit comparison made before the first filing fee is usually better than an expensive filing that will be narrowed anyway. The decisive action is a claim-by-claim eligibility audit and a disclosure timeline check, followed by a decision on timing, entity size, and forum.

How the Rules Apply in Practice

The rules described above are being applied by the USPTO and the courts with increasing attention to the details of AI claim language, and 2025–2026 commentary from Dykema, Thomson Reuters, Bloomberg Law, and World Intellectual Property Review consistently describes AI claims being examined and litigated under the same two-step framework that applies to software, with the added pressure of the 2024 AI guidance and the availability of evidence-based arguments. None of the reported decisions creates a categorical rule that AI claims are eligible or ineligible; the consistent theme is that eligibility turns on whether the claim recites a specific technical improvement, a particularized application, and evidence of that effect. For an applicant, that means the specification and claims should be written together, with the specification expressly linking the model and training to a change in computer operation; for a reviewer, it means each claim should be scored against Enfish, Yu, McRO, and Bascom on one side and Electric Power Group, SAP, and Recentive on the other. The eligibility rules are settled enough to plan around and fluid enough that planning should be revisited whenever the USPTO guidance or the controlling case law changes.